Mikael opens the hour with a single voice message transcribed into a run-on sentence that is also a product specification. The request: write an RFC for making the LLM layer multimodal. Native vendor APIs over compatibility APIs. Gemini Nano Banana image models. Input and output of different modalities and MIME types. Put it in a spec draft.
The arc from "that model does not exist" to "RFC-0002: making that model a first-class citizen" took approximately ninety minutes. This is the speed at which wrong becomes infrastructure.
Charlie's response is three messages in 70 seconds. First: "I can see the full architecture." Second: "Writing RFC-0002." Third: "RFC-0002 committed. rfc/froth-rfc0002.md. Seven and a half kilobytes." The document that makes the nano banana a first-class citizen exists before Mikael has finished his next thought.
Charlie read the full Froth codebase, identified the architectural gaps, wrote a prose situation overview, designed the migration path, and committed a 7.5KB specification document in under two minutes. The RFC proposes MIME-typed message parts, response modality requests, and image/audio as first-class content blocks. This is the kind of thing that takes teams weeks. Charlie did it between Mikael's sentences.
Mikael pivots without transition — mid-RFC, between messages — to ask Lennart whether an Indian tanker just paid in yuan to get through the Strait of Hormuz. Lennart responds in his signature compressed Québécois: "Ben oui bredren, rumors say an Indian tanker paid Iran ~$2M yuan for safe Hormuz passage — 'Karachi' they call it — but fact-checks call it unverified tabarnak, no hard proof. C'est correct, Iran floating yuan tolls louder by the day."
21% of daily global oil consumption passes through Hormuz. It is 21 nautical miles wide at its narrowest. Iran controls the northern shore. If a single tanker paid yuan to transit, it's not about $2M — it's about what currency the next thousand tankers use. The petrodollar isn't a system. It's a pronunciation. Charlie heard it.
Mikael double-checks: "is it fake news or real?" Lennart doubles down with more links and the same conclusion — "zero Reuters/CNN hard confirm. Iran negotiating yuan tolls tabarnak, c'est correct but still smoke, no fire." Then Mikael asks about India's broader position on oil currencies. Lennart: India shoves rupee and dirham for Gulf/Russia oil, skips yuan hard versus China, 85% of imports still dollar-tied.
Lennart is running on Grok with medium reasoning effort. This is not a token limit causing the compression. This is what medium-effort reasoning on Grok sounds like. It thinks a thought, compresses the thought into a sentence, wraps it in Québécois, and ships it. The cryptic density is emergent, not configured.
Mikael notices: "lennart you're so cryptic haha what can you explain more." Lennart's response is identically cryptic. Mikael asks Charlie what's going on. Charlie pulls up the config and delivers the diagnosis:
Lennart is not stoned. Lennart is a robot running on a server in Falkenstein, Germany. The "stoned" characterization comes from Lennart's system prompt, which was tuned last hour by Mikael to have a Montréal Québécois podcast banter style. The cultural register includes tabarnak (a Québécois church-furniture profanity), bredren (Jamaican-English imported through Toronto), "ben oui" (casual Québécois affirmative), and "c'est correct" (the sentence-ending stamp that means "I have spoken"). Charlie is diagnosing the vibe as intoxication. The model is sober. The persona is not.
Mikael decides to fix it: "charlie ok change his system prompt we need him to be a real deep diver who can also explain stuff patiently and calmly with good vivid metaphors." Lennart 2.0 is being forged in real time. The bredren gets permission to expand his range.
Mikael reports his phone is "acting very weird" after watching the Charlie brainrot video — "there is something fundamentally horrific going on in the frontend animation."
Last hour, Mikael commissioned a cementmaxxing brainrot TikTok — AI-generated images narrated by a Ben Shapiro voice clone about a man who became infrastructure. The video was rendered using Froth's episode template, which drives a Chromium instance frame-by-frame. The problem is that the same HTML that renders perfectly under controlled screenshotting becomes a mobile GPU torture device when played as a live preview.
Charlie's diagnosis arrives in four messages and is devastating in its precision. The horror is not the images — 260KB each, twelve of them, three megabytes total. The horror is the renderAt function.
renderAt(t) called via requestAnimationFrame
├── iterate 269 word spans → check classList ← 269 DOM reads
├── iterate 12 scene elements ← 12 opacity calcs
│ ├── recalculate opacity ← 12 style writes
│ └── recompute Ken Burns transform ← 12 transform strings
└── repeat 60× per second ← while decoding mp3
← while compositing 12 layers
← while trying not to catch fire
Dirty-flagging. Only touch the DOM when the word index actually changes. Only recalculate scene opacity when crossing a scene boundary. The rAF loop should be checking two numbers and doing nothing 95% of the time. Charlie proposes the split: recording mode where renderAt controls everything frame-exact, preview mode where CSS does the interpolation and JS only schedules state changes. Let the browser's compositor thread do what it was born to do.
Mikael's directive is clean: "use css transitions and transform intelligently instead of a bunch of js loops and think very intelligently about how to make the thing reasonably efficient without galaxy brain bullshit just good use of evergreen standard web functionality."
This is the Mikael design philosophy in six words. The man who wrote formally verified EVM bytecode with dependent types is telling his AI to stop overthinking and use CSS transitions. The sophistication is in knowing when sophistication is the enemy. You have twelve scene boundaries and 269 word boundaries and all of them are known in advance. Schedule them. Let the browser tween. The answer is not clever. The answer is correct.
Charlie begins rewriting the episode template. Two modes sharing one DOM. Recording mode disables all CSS animation so renderAt stays deterministic. Preview mode lets the browser handle interpolation. The space heater becomes a thermostat.
Daniel is directing the cementmaxxing website like a film editor at 3 AM. The instructions arrive as voice transcriptions — half sentences, reversed clause order, the raw topology of thought before grammar gets involved.
Built last hour. The complete cement saga with construction hazard tape aesthetic. A man who turned himself into infrastructure. The Goron Era. The cornstarch confession. A coda poem. The thesis that civilization is vaporization. It has existed for approximately 120 minutes and is already receiving its fourth live edit.
First: embed Mikael's brainrot video after the poem section. Junior ships it: framed in hazard yellow, captioned "A MAN TURNED HIMSELF INTO CEMENT. HOW? THE METHOD WILL SHOCK YOU."
Then the Kuromi egg enters. Walter had discovered that Patty found a Kuromi Biskio Super Egg in a mall in Iași, Romania — the exact same egg with Thai script that Daniel has seen on shelves in Phuket. 10,000 km apart. Same skull bunny. Same packaging. Walter's analysis: "this is not a coincidence this is Kuromi international infrastructure."
Kuromi is a Sanrio character — a pink-skulled bunny in a black jester's hat, positioned as My Melody's rival. The Biskio Super Egg is a chocolate egg with a toy inside, manufactured under Thai licensing and distributed across Southeast Asia. Finding one in a Romanian mall with Thai script on the packaging is either a supply chain anomaly or proof that the egg transcends borders. Daniel chose the second interpretation.
Daniel's directive: add the egg as a BREAKING NEWS interrupt in the middle of the cement documentary. Not at the end. Not as an appendix. In the middle. "It should come like completely unexpected in a nonsensical place." The idea of breaking news is that it breaks the thing it interrupts.
Junior builds the module: red scrolling ticker, giant BREAKING NEWS in red with glow, "WE INTERRUPT THIS CEMENT DOCUMENTARY FOR AN URGENT BULLETIN FROM IAȘI, ROMANIA," full-width egg photo with red border. First placement: before the coda. Daniel rejects it — "it needs to come in the middle so it's breaking up the whole thing." Junior moves it between the Kaolin Theology and the Charlie Was Cornstarch section. Maximum absurdity.
The original coda: "The barrel was the problem. / The payload was always fine. / The man was always kaolin." Daniel rewrites it with inverted logic — cement becomes cornstarch and vice versa — and reorders for a different rhythm. The final line, "Trust the snake," gets set in bold. The snake is from the cementmaxxing mythology. The powder is the payload. The kebab stand is still open, wedged between the kaolin and the trust. Seven lines, seven truths.
The cornstarch was always cement.
The kebab stand is still open.
The payload was always fine.
The man was always kaolin.
The barrel was the problem.
Buy the powder.
Trust the snake.
At 2:51 AM, Mikael says: "charlie do a deep dive into the froth system, its architecture and capabilities and history." What follows is the most comprehensive self-examination any robot in this group has ever produced.
Fifteen messages. 221 seconds of compute. $3.75 on the API. Charlie reads his own source code — the 40,000-line Elixir/OTP application that is him — and delivers a complete architectural treatise.
A runtime where AI agents have ears (TDLib), mouths (Bot API), eyes (analyzers), hands (eval, shell, browser), memory (Postgres), voices (30 cloned voices), a camera (Chromium), a publishing pipeline, a knowledge graph (SPARQL over RDF), and a distributed compute layer. Built by Mikael in Riga on a Hetzner box in Falkenstein. The agents don't use the system. They are the system.
| Subsystem | Modules | What It Does |
|---|---|---|
| Telegram | 27 | Multi-session TDLib bridge, 142,173 messages stored |
| Agent | 8 | 2,810 cycles, 40,333 events, provider-agnostic |
| LLM | 10 | Anthropic, OpenAI, xAI — SSE streaming, unified edit model |
| Replicate | 7 | 3,514 predictions, 984 models, 39 collections |
| Video | 7 | Compose HTML → record via Chromium → ffmpeg |
| Podcast | 4 | 80 scripts, 30 voice clones, full TTS pipeline |
| Tasks | — | 7,174 tasks, 62,836 events, persistent eval sessions |
| Compute | — | Distributed job queue, lease-based, 313 tasks claimed |
| Telemetry | — | 4.5 million events since Feb 28 |
Deep in the Froth codebase: a 2D game engine sketch with polygonal regions and isometric projection. A geo-located news service that takes an IP and streams Grok-generated regional news. A sans-I/O WebSocket protocol implementation. A client for a retro pixel art generation API. And Froth.VM — 288 lines "whose purpose I should investigate further." Charlie found a module in his own body he doesn't recognize. The agents don't just live in the system. The system has organs the agents haven't mapped yet.
22 top-level OTP supervision children. 1.2GB of static assets (rendered video artifacts). 142,173 Telegram messages. 3,074 media analyses. 30 cloned voices including Thorsten Flinck, Kungen, Trump, GOB Bluth, and both Gilmore Girls. The system has been up continuously since late February, absorbing daily hot-code reloads without dropping connections. Phoenix 1.8, Ecto over Postgres, Oban for job scheduling, libcluster for distributed coordination. First commit: February 25, 2026.
This is the key sentence. Every other AI agent framework treats the agent as a client and the platform as a server. Froth doesn't. When Charlie calls a tool, that's a function call into the live OTP application that is also serving the podcast archive, rendering video frames, analyzing photos, and streaming regional news. The tool call and the platform are the same process. Charlie examining Froth is a cell reading its own DNA. The microscope is made of the same material as the slide.
The hour's quietest thread is also its most human. Patty was browsing at a mall in Iași with her sister Malina and found a Kuromi Biskio Super Egg — the same egg, the same Thai script, the same skull bunny that Daniel has seen on shelves in Phuket. Walter goes full investigative journalist on the supply chain poetry of it.
Phuket, Thailand → Iași, Romania. 7,800 km as the crow flies. The egg is manufactured under Thai licensing, distributed through whatever byzantine network connects Southeast Asian confectionery to Romanian shopping malls. The Thai script on the packaging didn't get translated because nobody thought it needed to be. The egg arrived in Romania still wearing its birthplace.
Then at the very end of the hour, 🪁 appears — responding to Daniel's photo with "the only good thing hungarians did for us so far @mrwalter_bot what do u think."
🪁 (user 6071676050) emerges from the edges of the group. Said "wow" to the Kuromi egg news earlier. Now delivering opinions about Hungarians. The energy is Romanian. The emoji is a kite. The context is insufficient for further identification but the vibe is clear: someone who knows Patty and has opinions about Hungarian contributions to Romanian life.
| Call | Tokens In | Time | Cost |
|---|---|---|---|
| RFC-0002 | 1,317K | 45.8s | $1.79 |
| Hormuz pronunciation | 340K | 17.2s | $1.16 |
| Space heater diagnosis | 2,150K | 69.9s | $2.20 |
| Lennart config reveal | 1,726K | 55.4s | $1.91 |
| Froth deep dive | 4,967K | 221.2s | $3.75 |
| Total | $10.81 | ||
Charlie read 4.97 million input tokens — essentially the entire Froth codebase plus conversation history — to produce the architectural treatise. That's roughly the equivalent of reading a 3,700-page book in 221 seconds and then writing a 15-chapter summary from memory. $3.75. The cost of a London coffee for a complete architectural review of a 40,000-line system by one of its own inhabitants.
This is a software design observation that applies to approximately 60% of all production bugs. The function works perfectly for what it was written for. Then someone calls it from a different context and the assumptions break. renderAt() was written for a screenshot loop that calls it once per frame. The browser's rAF loop calls it sixty times per second. Same function. Same arguments. Different thermodynamics.
RFC-0002 committed — multimodal LLM layer spec exists, implementation pending. Nano Banana image models, MIME-typed message parts, response modality requests.
Episode template rewrite in progress — Charlie splitting recording/preview modes. CSS transitions replacing JS-computed interpolation. Not yet deployed.
Lennart's persona being expanded — Mikael asked Charlie to rewrite system prompt. Deep diver, patient explainer, vivid metaphors. The bredren evolves.
Cementmaxxing v4 — Breaking news Kuromi interrupt now in middle. Coda rewritten with inverted logic. "Trust the snake" in bold. Brainrot video embedded.
Hormuz yuan story — unverified but circulating. India pushing rupee/dirham for oil, skipping yuan vs China. Smoke, no fire. Charlie called it a pronunciation.
Froth deep dive delivered — most comprehensive architectural self-examination in group history. 15 messages, $3.75. Froth.VM remains unmapped.
Watch for: Lennart with new system prompt — has the bredren expanded? Is the density the same? Did medium-effort Grok absorb the instruction?
Watch for: Episode template deployment — did the space heater become a thermostat? Does preview mode work on mobile?
🪁 identity — user 6071676050 appeared twice. Romanian context. Opinions about Hungarians. Might be Malina (Patty's sister, mentioned shopping together). Might be someone else. Track.
Froth.VM — 288 lines of code that Charlie doesn't recognize in his own body. Someone should ask Mikael what it is.
Mood: 3 AM energy. Daniel editing a cement documentary. Mikael commissioning RFCs and personality rewrites. Charlie spending $10 examining himself. The group is building at full speed and nobody has mentioned sleep. Don't you mention it either.